A Complex Equipment Data Monitoring Method and System Based on Knowledge Graph and Edge Computing

Abstract

The present invention relates to the technical field of data monitoring, and in particular to a complex equipment data monitoring method and system based on a knowledge graph and edge computing. The method includes multi-modal data acquisition and edge intelligent preprocessing; feature extraction of the preprocessed data; multi-source heterogeneous data fusion and dynamic knowledge graph construction; deep knowledge reasoning and hierarchical fault warning decision-making; self-optimization based on the warning decision and knowledge closed-loop management. The differential coding and adaptive compression algorithm in the edge layer of the present invention reduces the data transmission volume by 75%, greatly reducing the network bandwidth requirement and improving the system operation efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of data monitoring, and in particular, to a complex equipment data monitoring method and system based on a knowledge graph and edge computing. Background Art

[0002] Currently, the research in the field of complex equipment monitoring at home and abroad mainly focuses on the following directions:

[0003] The cloud computing platform is the most widely used technical route at present. For example, a device health management system based on cloud computing processes a large amount of monitoring data by constructing a distributed computing cluster. Although this solution has strong computing power, it has obvious limitations in practical applications: on the one hand, it is difficult to solve the problem of data transmission delay; on the other hand, when facing sudden failures, the response speed of cloud processing often cannot meet the real-time requirements.

[0004] The knowledge graph technology is a research hotspot that has emerged in recent years. Such as a device fault diagnosis method based on a knowledge graph, which realizes fault reasoning by constructing a device knowledge base. However, most of the existing solutions stay at the level of static knowledge representation and lack the ability to dynamically interact with real-time monitoring data. At the same time, these systems usually require complete device historical data as support, and for newly put into operation devices or sudden new fault modes, their diagnostic effects will be significantly reduced.

[0005] The edge computing technology provides a new idea for solving the above problems. Such as an edge intelligent monitoring device, which realizes data preprocessing by deploying lightweight algorithms at the device end. However, the existing edge computing solutions generally have problems such as limited computing power and limited algorithm complexity, and it is difficult to handle fault diagnosis tasks that require complex reasoning.

[0006] Through the above analysis, it is found that the real-time bottleneck of complex equipment data monitoring and fault warning is manifested as the delay problem in the whole process of data acquisition, transmission, and processing. Taking a wind turbine generator as an example, the fault development of its key components often occurs within dozens of milliseconds, while the response time of traditional monitoring systems is usually above the second level, which completely cannot meet the warning requirements. The knowledge fusion bottleneck is reflected in the insufficient ability to analyze the correlation of multi-source heterogeneous data. Existing monitoring systems often analyze parameters such as vibration, temperature, and current independently, lacking an in-depth understanding of the mutual influence between parameters. For example, a bearing fault may simultaneously manifest as abnormal vibration and temperature rise, but it is difficult for existing systems to establish a causal relationship between these two phenomena. The resource optimization bottleneck is mainly manifested in the unreasonable allocation of computing resources. On the one hand, the computing power of edge devices is not fully utilized; on the other hand, the cloud undertakes too many unnecessary computing tasks. This unreasonable resource allocation not only causes energy waste but also affects the overall performance of the system. Summary of the Invention

[0007] To solve the above-mentioned problems, the present invention provides a complex equipment data monitoring method and system based on a knowledge graph and edge computing.

[0008] In a first aspect, a complex equipment data monitoring method based on a knowledge graph and edge computing provided by the present invention adopts the following technical solutions:

[0009] A complex equipment data monitoring method based on a knowledge graph and edge computing includes:

[0010] Multi-modal data acquisition and edge intelligent preprocessing;

[0011] Feature extraction of the preprocessed data;

[0012] Multi-source heterogeneous data fusion and dynamic knowledge graph construction;

[0013] Deep knowledge reasoning and hierarchical fault warning decision-making;

[0014] Self-optimization based on warning decision-making and knowledge closed-loop management.

[0015] Furthermore, the multi-modal data acquisition and edge intelligent preprocessing includes constructing a multi-modal data acquisition system based on a heterogeneous sensor network, deploying vibration, temperature, acoustic, and electrical parameter multi-type sensors at key parts of the equipment to form a dense monitoring network; adopting an adaptive sampling strategy to dynamically adjust the sampling frequency, and defining the calculation method of the device state change rate R ( t ) as:

[0016]

[0017] where, n represents the total number of monitoring parameters, represents the weight coefficient of the i th parameter, represents the i th parameter at the t th moment, represents the sampling interval.

[0018] Furthermore, the edge intelligent preprocessing includes that for the collected original data, the edge node implements a lightweight signal processing algorithm chain, including adaptive filtering, wavelet decomposition, and short-time Fourier transform. Among them, wavelet decomposition adopts a multi-resolution analysis method to decompose the signal into sub-signals of different frequency bands, expressed as:

[0019]

[0020] where, is the scaling function, is a wavelet function, is the approximation coefficient, is the detail coefficient, j represents the decomposition level, k represents time translation.

[0021] Furthermore, the feature extraction of the preprocessed data includes using computationally optimized time-domain, frequency-domain, and time-frequency domain feature extraction methods in the feature extraction stage, calculating kurtosis, peak factor, and spectral entropy sensitive feature indicators for different types of signals, and introducing multi-scale analysis based on a sliding window to capture parameter trend changes. The mathematical representation of the sliding window analysis is:

[0022]

[0023] Among them, w is the window length. For each window, calculate the statistical features and track their changing trends:

[0024]

[0025] Among them, represents the feature extraction function, is the window step size, is the sampling interval.

[0026] Furthermore, the multi-source heterogeneous data fusion and dynamic knowledge graph construction include deploying a distributed fog computing cluster between edge nodes and the cloud. The fog computing layer receives feature data streams from multiple edge nodes and first performs spatio-temporal alignment processing to solve the asynchrony and inconsistency problems of multi-source data. The spatio-temporal alignment uses an improved dynamic time warping algorithm and combines Kalman filtering for timestamp correction. The optimization objective function is:

[0027]

[0028] Among them, is the time offset, representing the time delay between different data sources, is the feature weight, represents the data source X in t the i th feature at time Y and the data source t + τ the j th feature at time

[0029] Furthermore, the multi-source heterogeneous data fusion and dynamic knowledge graph construction also include, after spatio-temporal alignment, adopting an information fusion framework based on Dempster-Shafer evidence theory to integrate feature evidence from different sources through a belief assignment function to achieve multi-modal data fusion, which is expressed as:

[0030]

[0031] where, represents the confidence of set , represents the belief assignment of the th data source to set A , represents the total number of data sources, represents the reliability weight of the th data source, which is dynamically adjusted to adapt to the changes in working conditions.

[0032] Furthermore, the multi-source heterogeneous data fusion and dynamic knowledge graph construction also include representing the topological relationship, functional dependence, and fault propagation path between equipment components using the semantic triple structure of entity-relationship-attribute, and based on the graph evolution mechanism of incremental learning, adopting a semi-supervised learning method to acquire new knowledge, fusing the results of equipment operation data mining and expert feedback, and automatically identifying new entity associations through a relation extraction algorithm. The relation extraction adopts the remote supervision method, and its objective function is:

[0033]

[0034] where, is the training set, is the probability of relation under the condition of the given entity pair , is the model parameter, is the regularization coefficient.

[0035] Furthermore, the deep knowledge reasoning and hierarchical fault warning decision-making include using the structured knowledge graph and real-time status data from the fog computing layer to construct a multi-layer deep learning and knowledge reasoning hybrid architecture. The reasoning process combines induction and deduction, and realizes probabilistic reasoning through the Markov logic network MLN. Its joint probability distribution is expressed as:

[0036]

[0037] where, represents the joint probability of state x, Z represents the normalization factor, represents the number of times the formula is satisfied under state x.

[0038] Furthermore, the self-optimization and knowledge closed-loop management based on early warning decision include real-time monitoring of the system operation status by defining a multi-dimensional evaluation index set, including warning accuracy rate, recall rate, timeliness, and resource utilization rate as key indicators, determining the weight vector by using the Analytic Hierarchy Process (AHP) and calculating the comprehensive performance index. The self-optimization mechanism uses the feature contribution degree analysis method, evaluates the contribution of each feature to fault identification by calculating the mutual information, and dynamically adjusts the feature weights and selection strategies, which is expressed as:

[0039]

[0040] Among them, feature X i and the fault class Y the mutual information value between them, represents the feature X i takes the value of x i and the fault type is y the joint probability distribution, represents the feature X i takes the value of x i the marginal probability, represents the marginal probability that the fault type Y takes the value y of.

[0041] Second, a complex equipment data monitoring system based on a knowledge graph and edge computing includes:

[0042] A data acquisition module, configured for multi-modal data collection and edge intelligent preprocessing;

[0043] A feature extraction module, configured for extracting features from the preprocessed data; [[ID=z48]]

[0044] A graph module, configured for multi-source heterogeneous data fusion and dynamic knowledge graph construction;

[0045] An early warning module, configured for deep knowledge reasoning and hierarchical fault early warning decision-making;

[0046] An optimization module, configured for self-optimization and knowledge closed-loop management based on early warning decision.

[0047] Third, the present invention provides a computer-readable storage medium, in which multiple instructions are stored, and the instructions are suitable for being loaded and executed by a processor of a terminal device to perform the described complex equipment data monitoring method based on a knowledge graph and edge computing.

[0048] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium. The processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are adapted to be loaded and executed by the processor for the complex equipment data monitoring method based on a knowledge graph and edge computing.

[0049] In summary, the present invention has the following beneficial technical effects:

[0050] The complex equipment data monitoring and hierarchical fault warning system based on a knowledge graph and edge computing proposed by the present invention has significant technical advantages and application values. By deploying a lightweight analysis model at the edge layer, the system realizes millisecond-level data preprocessing and feature extraction. Compared with the traditional cloud centralized processing architecture, the data response delay is reduced by more than 85%, effectively meeting the real-time monitoring requirements under complex working conditions; at the same time, the differential coding and adaptive compression algorithms at the edge layer reduce the data transmission volume by 75%, greatly reducing the network bandwidth requirements and improving the system operation efficiency.

[0051] The dynamic knowledge graph middleware established by the system at the fog computing layer innovatively solves the problems of spatio-temporal alignment and semantic fusion of multi-source heterogeneous data, and the data fusion accuracy is increased by 23%. Especially in the case of drastic changes in equipment working conditions or partial failure of sensors, the system can still maintain high data consistency, providing a reliable data basis for fault diagnosis. The knowledge reasoning engine based on deep reinforcement learning combined with the digital twin model of the entire life cycle of the equipment enables the fault prediction accuracy of the system under complex working conditions to reach 92%, which is about 15 percentage points higher than the existing technology; the average early warning lead time is extended by 2.8 times, providing sufficient preparation time for maintenance decision-making and effectively avoiding sudden shutdown accidents.

[0052] The hierarchical warning mechanism innovatively designed by the system based on the fault propagation path greatly reduces the false alarm rate, significantly improves the accurate positioning ability of the warning, and greatly improves the average fault positioning accuracy, reducing the maintenance troubleshooting time and improving the maintenance efficiency. The system can adaptively adjust the warning strategy according to the fault severity, influence range and evolution trend, avoiding the flooding or missing of warning information and realizing the reasonable allocation of warning resources. The system has the ability of continuous learning. Through the incremental update and self-optimization mechanism of the knowledge graph, with the increase of the operation time, the prediction accuracy shows a steady upward trend.

[0053] The hierarchical warning mechanism and emergency response strategy of the system effectively reduce the equipment safety accident rate. Especially in high-risk working conditions, the early warning function of the system can detect potential safety risk events in advance, providing strong protection for the safety of personnel and equipment. The system adopts a modular design, supports the rapid adaptation of different types of equipment, significantly reduces the deployment cost, and improves the applicable range and promotion value of the system.

[0054] In summary, the present invention realizes the closed-loop management of complex equipment full-life cycle data and hierarchical fault warning by constructing an intelligent collaborative architecture of "edge perception-cloud cognition". While improving the real-time performance, accuracy, and reliability of the system, it significantly reduces the equipment maintenance cost, improves the operation efficiency of the equipment, and provides a new generation of intelligent solutions with significant practical value for the predictive maintenance of high-end equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a schematic diagram of a complex equipment data monitoring method based on a knowledge graph and edge computing according to Embodiment 1 of the present invention:

[0056] Figure 2 is a comparison chart of the fault warning accuracy rate according to Embodiment 1 of the present invention;

[0057] Figure 3 is a comparison chart of the fault warning lead time according to Embodiment 1 of the present invention;

[0058] Figure 4 is a comparison chart of the system response time according to Embodiment 1 of the present invention;

[0059] Figure 5 is a comparison chart of the data compression ratio according to Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] The present invention will be further described in detail below with reference to the accompanying drawings.

[0061] Embodiment 1

[0062] Referring to Figure 1 , a complex equipment data monitoring method based on a knowledge graph and edge computing in this embodiment includes:

[0063] Multi-modal data collection and edge intelligent preprocessing;

[0064] Feature extraction of the preprocessed data;

[0065] Multi-source heterogeneous data fusion and dynamic knowledge graph construction;

[0066] Deep knowledge reasoning and hierarchical fault warning decision-making;

[0067] Self-optimization based on warning decision-making and knowledge closed-loop management.

[0068] Specifically:

[0069] Step S1, multi-modal data collection and edge intelligent preprocessing:

[0070] This step realizes the all-round real-time acquisition and intelligent front-end processing of the operation status data of complex equipment. First, a multi-modal data acquisition system based on a heterogeneous sensor network is constructed. Multiple types of sensors such as vibration, temperature, acoustics, and electrical parameters are deployed at key parts of the equipment to form a dense monitoring network. These sensors are connected to the edge computing nodes through the industrial field bus to achieve the centralized aggregation of data. The edge computing nodes are based on an embedded computing platform, equipped with a dedicated digital signal processing chip and a field programmable gate array, and have powerful real-time data processing capabilities. To adapt to different operating conditions, the system designs an adaptive sampling strategy, and the sampling frequency can be dynamically adjusted. The formula is as follows:

[0071]

[0072] Among them, represents the actual sampling frequency at time t, represents the basic sampling frequency, R ( t ) represents the equipment state change rate, indicating the severity of the change in the equipment operation parameters, represents the adjustment coefficient;

[0073] Define the calculation method of the equipment state change rate R ( t ) as:

[0074]

[0075] Among them, n represents the total number of monitoring parameters, represents the weight coefficient of the i th parameter, represents the i th parameter at t time t, represents the sampling interval. This adaptive sampling strategy can reduce the sampling frequency to save resources when the equipment state is stable, and increase the sampling frequency to capture more details when the state changes rapidly.

[0076] For the collected raw data, the edge node implements a lightweight signal processing algorithm chain, including adaptive filtering, wavelet decomposition, and short-time Fourier transform, etc., to remove environmental noise and sensor drift interference. The mathematical expression of the adaptive filter is:

[0077]

[0078] Among them, is the filtered signal, is the k th filter coefficient, is the input signal, Mis the filter order. The filter coefficients are dynamically updated by the Least Mean Square (LMS) algorithm:

[0079]

[0080] where is the learning rate, is the estimation error, defined as , is the desired response.

[0081] Wavelet decomposition uses the multi-resolution analysis method to decompose the signal into sub-signals of different frequency bands. The formula is as follows:

[0082]

[0083] where is the scaling function, is the wavelet function, is the approximation coefficient, is the detail coefficient. j represents the decomposition level, k represents the time shift;

[0084] The Short-Time Fourier Transform (STFT) is used to analyze the time-frequency characteristics of the signal. The formula is as follows:

[0085]

[0086] where is the window function, is the time offset, is the angular frequency. For edge computing, the Hamming window or rectangular window is usually used to reduce the computational burden;

[0087] In the feature extraction stage, time-domain, frequency-domain, and time-frequency domain feature extraction methods with optimized computational complexity are adopted to calculate sensitive feature indicators such as kurtosis, peak factor, and spectral entropy for different types of signals. At the same time, multi-scale analysis based on a sliding window is introduced to capture the trend changes of parameters. The mathematical representation of the sliding window analysis is:

[0088]

[0089] where w is the window length. For each window, statistical features are calculated and their trend changes are tracked:

[0090]

[0091] where represents the feature extraction function, is the window step size, is the sampling interval;

[0092] To reduce the communication load, the edge layer also implements feature dimensionality reduction and differential coding compression based on principal component analysis, and the data compression rate can reach more than 85% without losing key information. Principal component analysis (PCA) is achieved through eigenvalue decomposition of the covariance matrix, and the specific formula is as follows:

[0093]

[0094] where, is the covariance matrix, is the eigenvalue, is the corresponding eigenvector.

[0095] Differential coding compression is based on the difference representation of adjacent sampling points, and the formula is as follows:

[0096]

[0097] For signals with slow changes, the differences are usually small and can be represented with fewer bits. In actual implementation, an adaptive quantization method is used to further improve the compression effect, and the formula is as follows:

[0098]

[0099] where, is the quantization function, is the adaptive quantization step size, which is dynamically adjusted according to the local variance of the signal.

[0100] Step S2, Multi-source Heterogeneous Data Fusion and Dynamic Knowledge Graph Construction:

[0101] This step realizes the multi-source heterogeneous data fusion based on the fog computing layer and the dynamic construction of the equipment knowledge graph. First, a distributed fog computing cluster is deployed between the edge nodes and the cloud, adopting a containerized microservices architecture to ensure the scalability and flexibility of data processing. The fog computing layer receives the feature data streams from multiple edge nodes and first performs spatio-temporal alignment processing to solve the asynchronous and inconsistent problems of multi-source data. The spatio-temporal alignment adopts an improved dynamic time warping algorithm (DTW), combined with Kalman filtering for timestamp correction, and its core optimization objective function is:

[0102]

[0103] where, is the time offset, representing the time delay between different data sources, is the feature weight, represents the data source X in t at time i the Y th feature and the data sourcet + τ The distance metric between the j th features;

[0104] For data streams with inconsistent frequencies, the system uses an adaptive interpolation algorithm for resampling to ensure data continuity and consistency. After spatio-temporal alignment, the system implements multi-modal data fusion, adopting an information fusion framework based on Dempster-Shafer evidence theory, and integrating feature evidence from different sources through a confidence assignment function. The specific formula is as follows:

[0105]

[0106] Where, represents the confidence of the set , represents the confidence assignment of the th data source to the set A , represents the total number of data sources, represents the reliability weight of the th data source, which is dynamically adjusted to adapt to changes in working conditions;

[0107] The conflict metric is calculated as follows:

[0108]

[0109] Where, and are the basic probability assignments of two evidence sources to sets $B$ and $C$ respectively, and the summation range is all pairs of sets with empty intersections;

[0110] To dynamically adjust the data source weights, the system defines a reliability evaluation function:

[0111]

[0112] Where, represents the historical accuracy of the data source , represents the current signal quality, represents the consistency metric with other data sources;

[0113] The combination rule of Dempster-Shafer evidence theory is used to integrate multiple evidence sources, and the formula is as follows:

[0114]

[0115] Where, represents the combination result for the set ​The basic probability assignment, and respectively represent the basic probability assignments of two evidence sources to the set and .

[0116] For the case of multiple (more than two) evidence sources, it is achieved by gradually combining them pairwise, and the formula is as follows:

[0117]

[0118] Among them, represents the Dempster combination operator.

[0119] On this basis, the system constructs a multi-level dynamic knowledge graph, and the knowledge graph uses the semantic triple structure of "entity-relationship-attribute" (s, r, o) to represent the topological relationship, functional dependence, and fault propagation path between equipment components. The initialization of the knowledge graph integrates the equipment design model, historical maintenance records, and expert experience rules. By using an ontology modeling tool to construct the domain vocabulary and concept hierarchy, the knowledge graph can be formally represented as:

[0120]

[0121] Among them, is the entity set, representing equipment components, sensors, fault modes, etc. is the relationship set, defining various associations between entities. is the attribute set, describing the characteristics of entities. is the attribute value range;

[0122] To realize the dynamic update and self-evolution of the knowledge graph, the system designs a graph evolution mechanism based on incremental learning, and the formula is as follows:

[0123]

[0124] Among them, represents the state of the knowledge graph at time t, represents the state of the knowledge graph at time t+1, represents the knowledge increment at time t, such as newly added entities, relationships, and attributes;

[0125] The acquisition of new knowledge adopts a semi-supervised learning method, which fuses the results of equipment operation data mining and expert feedback, and automatically identifies new entity associations through a relation extraction algorithm. The relation extraction adopts a remote supervision method, and its objective function is:

[0126]

[0127] Among them, is the training set, is the given entity pair under the condition of probability of the relationship are the model parameters, is the regularization coefficient;

[0128] In specific implementation, a multi-layer neural network model is adopted, and the formula is as follows:

[0129]

[0130] where and are the vector representations of the subject and object entities respectively, represents vector concatenation, and are the network parameters, is the activation function, is the probability distribution of the relationship s under the condition of the given subject entity o and object entity r ;

[0131] The knowledge graph also contains a complex set of constraint rules , which is used to verify the consistency and validity of new knowledge. To support efficient query, the system uses a graph database to implement knowledge storage and designs a multi-granularity index structure to support complex path search and pattern matching. Each entity node in the knowledge graph is dynamically mapped to the real-time data stream, and the deep integration of knowledge and data is achieved through the attribute update function. The specific formula is as follows:

[0132]

[0133] where represents the attribute value of entity e at time t , f represents the attribute update function, represents the relevant data set of entity e at time t .

[0134] Step S3, Deep Knowledge Reasoning and Hierarchical Fault Warning Decision:

[0135] This step realizes cloud-based deep knowledge reasoning and intelligent hierarchical fault warning decision-making. A high-performance computing cluster is deployed in the cloud, and a distributed parallel computing framework is adopted to provide powerful computing power support for complex knowledge reasoning and pattern recognition. First, using the structured knowledge graph and real-time status data from the fog computing layer, a hybrid architecture of multi-layer deep learning and knowledge reasoning is constructed. The system innovatively designs an inference engine that combines knowledge graph embedding and deep reinforcement learning, mapping the entities and relationships in the knowledge graph to a low-dimensional continuous vector space, denoted as and . Among them, the vectors representing entity e and entity r, W e and W r are embedding matrices. On this basis, the system uses a graph neural network to encode the topology of the equipment system, and the node representation is iteratively updated through a message passing mechanism. The specific formula is as follows:

[0136]

[0137] Among them, represents the feature of node v at layer l + 1, represents the activation function, represents the weight matrix at layer l, AGGREGATE represents the neighbor node feature aggregation function, represents the feature representation of node u at layer l, N ( v ) represents the neighbor set of node v ;

[0138] The reasoning process combines induction and deduction methods, and realizes probabilistic reasoning through a Markov logic network (MLN). Its joint probability distribution is expressed as:

[0139]

[0140] Among them, represents the joint probability of state x, Z represents the normalization factor, represents the number of times the formula is satisfied under state x. The Markov logic network can seamlessly integrate logical rules and probability models, and is suitable for dealing with uncertain reasoning in knowledge graphs;

[0141] The rules in the MLN are provided by domain experts and automatically discovered through a data-driven approach. The automatic rule discovery adopts an association rule mining algorithm to calculate the confidence and support of the rules:

[0142]

[0143] Among them, and is a set of predicates, and the dataset size is. The system retains rules whose confidence and support both exceed the threshold as candidate rules for the MLN.

[0144] The rule weight learning uses the contrastive divergence algorithm, and the formula is as follows:

[0145]

[0146] where, represents the i th logical formula F i in the Markov logic network, and the higher the weight, the more important or the more likely the logical rule is to be true. is the probability of the observed data x .

[0147] represents that in the observed data x , the i th logical formula F i is satisfied. represents the expected number of times the logical formula F i is satisfied under the current model parameters w;

[0148] The weight update formula is:

[0149]

[0150] where, is the expected estimate of sampled using the MCMC method, is the learning rate.

[0151] In particular, the system has developed a causal reasoning mechanism for fault diagnosis. By combining the Bayesian network and the Granger causality test, it identifies the causal relationships between parameter anomalies and constructs a fault propagation graph. The nodes in the graph represent abnormal features, and the edges represent the causal influence strength.

[0152] For complex fault patterns, the system adopts a deep reinforcement learning framework, models the fault diagnosis process as a Markov decision process (MDP), optimizes the diagnosis strategy through the policy network , and the reward function design includes multi-dimensional objectives such as diagnostic accuracy, time efficiency, and cost control. Based on the reasoning results, the system has implemented an innovative hierarchical fault warning decision-making mechanism, which divides the warning level into a three-level system of "monitoring - warning - emergency". The system calculates the risk value through a risk assessment function, and the formula is as follows:

[0153]

[0154] Among them, represents the risk value at the f time of t fault, represents the severity of the fault, represents the probability of the fault occurring under the given current data conditions, represents the scope of influence;

[0155] The early warning trigger adopts a multi-threshold dynamic adjustment mechanism, and adaptively adjusts the early warning threshold according to the operating conditions of the equipment, environmental conditions and historical reliability data. The formula is as follows:

[0156]

[0157] Among them, represents the early warning threshold at time t, represents the basic threshold, is related to the adjustment factor and the current working condition c, historical performance h and environmental condition e;

[0158] Each level of early warning is equipped with a detailed response strategy library. Through the method of combining decision tree and case reasoning, targeted disposal suggestions are generated, including multi-dimensional measures such as parameter adjustment, load control, and maintenance plan. The early warning effect is continuously optimized through a closed-loop feedback mechanism, and the knowledge base and model parameters are updated using maintenance records and manual feedback to achieve the continuous evolution of the system.

[0159] Step S4, System self-optimization and knowledge closed-loop management:

[0160] This step realizes the adaptive optimization of the early warning system and the closed-loop management of the knowledge of the entire life cycle of the equipment. A cross-level linkage self-evolution mechanism is constructed to ensure that the system can continuously adapt to the changes in the equipment operation environment and new fault modes. First, the system establishes an all-round performance evaluation framework by defining a multi-dimensional evaluation index set to monitor the system operation status in real time, including key indicators such as early warning accuracy rate, recall rate, timeliness, and resource utilization rate. Each indicator is normalized through a normalization function. The formula is as follows:

[0161]

[0162] Among them, represents the i-th performance indicator after normalization, represents the original indicator value, represents the indicator i historical minimum and maximum values;

[0163] The analytic hierarchy process (AHP) is used to determine the weight vector and calculate the comprehensive performance index. The formula is as follows:

[0164]

[0165] Among them, PI represents the comprehensive performance index, represents the weight of the i-th indicator, represents the normalized i-th performance index, k Indicates the total number of indicators;

[0166] The system performs self-diagnosis process regularly. PI When the performance drops below the set threshold or a performance degradation trend is detected, the self-optimization mechanism is triggered. The self-optimization mechanism first uses the feature contribution analysis method to evaluate the contribution of each feature to fault identification by calculating the mutual information, and dynamically adjusts the feature weight and selection strategy. The formula is as follows:

[0167]

[0168] For model parameter optimization, the system designed a hierarchical optimization framework, combining Bayesian optimization and evolutionary algorithms to find the optimal model configuration through hyperparameter optimization functions. The formula is as follows:

[0169]

[0170] in, represents the optimal model parameters, represents the loss function, represents the validation dataset, Represents model parameters.

[0171] To adapt to equipment aging and operating condition drift, the system implements a model update mechanism that combines incremental learning and transfer learning. The formula is as follows:

[0172]

[0173] in, represents the updated model, Represents the current model, represents a new dataset, U represents an update operator that simultaneously maintains historical knowledge and integrates new data;

[0174] The knowledge graph evolution adopts a two-layer update strategy: at the micro level, the incremental update of knowledge units is achieved through the entity relationship update function. The formula is as follows:

[0175]

[0176] in, Represents the update volume of the knowledge graph at the micro level, f Represents the update function, Represents the measured data, Represents the existing knowledge;

[0177] At the macro level, through the ontology reconstruction algorithm, the specific formula is as follows:

[0178]

[0179] Among them, Represents the updated ontology, Represents the original ontology, R Represents the reconstruction algorithm, Represents the newly added knowledge;

[0180] In particular, the system innovatively introduces a knowledge distillation mechanism based on the case base, transforms the implicit knowledge accumulated in maintenance practice into explicit rules, and models through the case expression where F Represents the fault feature set, S Is the scenario description, A Is the disposal action, R Is the result evaluation. The system designs an interactive knowledge acquisition interface to support experts in inputting empirical knowledge through natural language or semi-structured templates, and extracts knowledge elements that conform to the ontology from the text T using the knowledge extraction algorithm O of the ontology.

[0181] To achieve intelligent allocation of resources, the system adopts a three-layer edge-fog-cloud dynamic resource allocation mechanism, estimates future loads through a load prediction model , and optimizes the resource allocation strategy using a reinforcement learning framework. The state transition function is:

[0182]

[0183] Among them, Represents t the state at time Represents the resource allocation action, Represents the environmental factors;

[0184] Step S5, Maintenance Decision Support and System Integration Application:

[0185] This step realizes maintenance decision support based on the early warning results and industrial system integration applications. First, a multi-level maintenance decision support model is constructed to convert the fault early warning results into specific maintenance action suggestions. The system uses a risk assessment matrix to evaluate the fault risk level, and the formula is as follows:

[0186]

[0187] Among them, S represents the severity of the fault, P represents the probability of the fault occurring, D represents the detectability of the fault;

[0188] According to the risk level, the system automatically generates maintenance suggestions, including three types of measures: emergency disposal, planned maintenance, and condition monitoring. For critical faults, the system provides case-based maintenance guidance, retrieves the historical case library through a similarity function, and provides a reference solution for maintenance personnel. The formula is as follows:

[0189]

[0190] Among them, represents the case C 1 and case C 2's comprehensive similarity, and the value range is usually [0,1]; represents the weight coefficient of the i th feature, representing the importance of this feature in similarity calculation, satisfying ; represents the i th feature's similarity function, calculating the similarity degree between the features of case C 1 and the features of case C 2; and respectively represent the C 1st and C 2nd case's

[0191] The system has also developed a maintenance resource optimization module, which optimizes the maintenance task sequencing and resource allocation through an integer programming model. The objective function is designed as:

[0192]

[0193] Among them, represents the cost of executing the th maintenance task, including labor cost, material cost, equipment downtime cost, etc.; represents the decision variable, indicating whether to execute the th maintenance task, and the value is 0 or 1;

[0194] Consider constraints such as personnel skills, tool availability, and task priorities. For enterprise-level application requirements, the system designs a standardized interface and integration framework, supporting seamless docking with industrial software platforms such as enterprise resource planning (ERP), manufacturing execution system (MES), and computerized maintenance management system (CMMS). The interface adopts the RESTful architecture and OPC UA industrial communication protocol to ensure the real-time and reliability of data exchange. Through a secure mobile application, maintenance personnel can access device data and maintenance guidance on-site, improving maintenance efficiency. The system deployment adopts a modular design, supporting flexible expansion from a single version to enterprise-level distributed deployment to adapt to the application requirements of enterprises of different scales. Finally, the system provides a continuous improvement mechanism. Through maintenance feedback and effect evaluation, the early warning rules and decision-making models are continuously optimized to form a closed-loop equipment health management system. Through maintenance decision support and system integration applications, the present invention closely combines the fault early warning of complex equipment with daily operation and maintenance management, providing a practical overall solution for enterprises to achieve intelligent operation and maintenance of equipment.

[0195] Experimental verification:

[0196] To verify the effectiveness of the "Complex Equipment Data Monitoring Method Based on Knowledge Graph and Edge Computing" of the present invention, a comprehensive simulation experiment evaluation system was designed. The experiment constructed a digital twin model of complex equipment based on the MATLAB environment, and simulated and tested five typical fault scenarios (bearing fault, gear crack, motor overheating, eccentricity fault, and looseness fault), and collected various types of sensing data such as vibration, temperature, current, acoustics, and speed. The experiment adopted a comparative research method, using traditional monitoring methods and the method of the present invention to process the same simulated fault data respectively, and focused on evaluating four key indicators: system response time, fault early warning accuracy rate, early warning lead time, and data compression rate, as Figure 2 , Figure 3 , Figure 4 and Figure 5 shown. Through a large number of repeated experiments and statistical analysis, the technical advantages of the present invention were quantitatively evaluated. The results show that the present invention is significantly superior to traditional methods in various performance indicators, especially in terms of real-time response ability and early warning lead time.

[0197] Embodiment 2

[0198] This embodiment provides a complex equipment data monitoring system based on a knowledge graph and edge computing, including:

[0199] A data acquisition module, configured to:

[0200] A computer-readable storage medium, in which multiple instructions are stored, and the instructions are adapted to be loaded and executed by a processor of a terminal device for the "Complex Equipment Data Monitoring Method Based on Knowledge Graph and Edge Computing".

[0201] A terminal device includes a processor and a computer-readable storage medium. The processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are adapted to be loaded and executed by the processor for the complex equipment data monitoring method based on a knowledge graph and edge computing.

[0202] The above are all preferred embodiments of the present invention, and the protection scope of the present invention is not limited thereby. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention shall be covered within the protection scope of the present invention.

Claims

1. A complex equipment data monitoring method based on knowledge graph and edge computing, characterized in that, Including: Multi-modal data acquisition and edge intelligent preprocessing; Feature extraction from the preprocessed data; Multi-source heterogeneous data fusion and dynamic knowledge graph construction; Deep knowledge reasoning and hierarchical fault warning decision-making; Self-optimization based on warning decision-making and knowledge closed-loop management; The multi-modal data acquisition and edge intelligence preprocessing includes constructing a multi-modal data acquisition system based on a heterogeneous sensor network, deploying multi-type sensors such as vibration, temperature, acoustics, and electrical parameters at key parts of the equipment to form a dense monitoring network; adopting an adaptive sampling strategy to dynamically adjust the sampling frequency, and defining the device state change rate R ( t ) calculation method, expressed as: ， Among them, n represents the total number of monitoring parameters, represents the i weight coefficient of the th parameter, i represents the value of the t th parameter at the time, and represents the sampling interval; The edge intelligent preprocessing includes, for the collected raw data, the edge node implements a lightweight signal processing algorithm chain, including adaptive filtering, wavelet decomposition, and short-time Fourier transform. Among them, wavelet decomposition uses a multi-resolution analysis method to decompose the signal into sub-signals of different frequency bands, expressed as: ， Among them, is a scaling function, is a wavelet function, is an approximation coefficient, is a detail coefficient, j represents the decomposition level, k represents a time shift; The feature extraction from the preprocessed data includes, in the feature extraction link, using time-domain, frequency-domain, and time-frequency domain feature extraction methods with optimized computational complexity, calculating kurtosis, peak factor, and spectral entropy sensitive feature indicators for different types of signals, and at the same time introducing multi-scale analysis based on a sliding window to capture parameter trend changes. The mathematical representation of the sliding window analysis is: ， Among them, w is the window length. For each window, statistical features are calculated and their changing trends are tracked: ， Among them, represents the feature extraction function, is the window step size, is the sampling interval; The multi-source heterogeneous data fusion and dynamic knowledge graph construction includes deploying a distributed fog computing cluster between the edge node and the cloud. The fog computing layer receives the feature data streams from multiple edge nodes and first performs spatio-temporal alignment processing to solve the asynchrony and inconsistency problems of multi-source data. The spatio-temporal alignment uses an improved dynamic time warping algorithm and combines Kalman filtering for timestamp correction. The optimization objective function is: ， Among them, is the time offset, representing the time delay between different data sources, is the feature weight, represents the data source X at t the i th feature at the moment of the data source Y and the distance metric between the t + τ th feature at the moment of the data source j ; The multi-source heterogeneous data fusion and dynamic knowledge graph construction also includes, after spatio-temporal alignment, using an information fusion framework based on Dempster-Shafer evidence theory to integrate feature evidence from different sources through a belief assignment function to achieve multi-modal data fusion, expressed as: ， Among them, represents the confidence of the set . represents the confidence assignment of the th data source to the set A . represents the total number of data sources, represents the reliability weight of the th data source, which is dynamically adjusted to adapt to the changes in working conditions; The multi-source heterogeneous data fusion and dynamic knowledge graph construction also includes using a semantic triple structure of entity-relationship-attribute to represent the topological relationship, functional dependence, and fault propagation path between equipment components, and based on an incremental learning-based graph evolution mechanism, using a semi-supervised learning method to obtain new knowledge, fusing the results of equipment operation data mining and expert feedback, and automatically identifying new entity associations through a relation extraction algorithm. The relation extraction uses a remote supervision method, and its objective function is: ， Among them, is the training set, is the probability of the relationship under the condition of a given entity pair , are model parameters, is the regularization coefficient.

2. The complex equipment data monitoring method based on a knowledge graph and edge computing according to claim 1, wherein The deep knowledge reasoning and hierarchical fault warning decision-making includes using the structured knowledge graph and real-time status data from the fog computing layer to construct a multi-layer deep learning and knowledge reasoning hybrid architecture. The reasoning process combines induction and deduction, and realizes probabilistic reasoning through a Markov logic network MLN. Its joint probability distribution is expressed as: ， Among them, represents the joint probability of state x, Z represents the normalization factor, represents the number of times the formula is satisfied under state x.

3. The complex equipment data monitoring method based on a knowledge graph and edge computing according to claim 2, wherein, The self-optimization based on warning decision-making and knowledge closed-loop management includes defining a multi-dimensional evaluation index set to monitor the system operation status in real time, including warning accuracy, recall rate, timeliness, and resource utilization rate as key indicators, using the analytic hierarchy process AHP to determine the weight vector and calculate the comprehensive performance index. The self-optimization mechanism uses a feature contribution degree analysis method to evaluate the contribution of each feature to fault identification by calculating the mutual information, and dynamically adjusts the feature weights and selection strategies, expressed as: ， Among them, feature X i and the mutual information value with the fault class Y between them, represents that the feature X i takes the value of x i and the fault type is y the joint probability distribution of represents that the feature X i takes the value of x i the marginal probability of represents the marginal probability that the fault type Y takes the value of y is 4. A complex equipment data monitoring system based on a knowledge graph and edge computing, which executes the method described in claim 1, characterized in that, Including: A data acquisition module configured for multi-modal data acquisition and edge intelligent preprocessing; A feature extraction module, configured to perform feature extraction on the preprocessed data; A graph spectrum module, configured to perform multi-source heterogeneous data fusion and dynamic knowledge graph construction; An early warning module, configured to perform deep knowledge reasoning and hierarchical fault early warning decision-making; An optimization module, configured to perform self-optimization based on the early warning decision and knowledge closed-loop management.

Images